Skip to content

Speech benchmarks

How often does a model miss a word? Compare transcription scores across open models and speech APIs. Select a point to see where its number came from.

Public checkpoint measurements

Both authors ran the Open ASR harness on an H200. The runs used different sets of English recordings. Neither run included the leaderboard’s two private tracks.

CheckpointMean WERAggregate RTFxEvaluation
OrukQwen-ASR-1.7B4.76%718.84×Measured · Nine public English scoring columns, including both LibriSpeech splits. The additional URGENT2024-Clean split is excluded from the mean. Run of 2026-11-06.
Oruk Fourier-2B4.95%267.80×Measured · Seven default public English short-form splits. The additional URGENT2024-Clean split is excluded from the mean.

Download all ASR rows with evidence status, protocol and source in CSV or JSON.

Open ASR Leaderboard

Public + private English sets
Open ASR Leaderboard · model size and word error rateModel-family size in billions on a log scale. Closed APIs have a separate lane. Lower word error rate is better. The star marks Oruk v1, 5.97% WER, board snapshot. Model details are available in the selector below.WER (%) ↓Open ASR · English full boardCLOSED APIs4.55.05.56.06.57.07.58.08.59.00.5B1B2B5B10B30BNominal model size · log scaleZoom Scribe v2.1 Pro · 4.87% · Board snapshotZoom ScribeOtter Speech v3.1 · 5.21% · Board snapshotAzure Speech 09-2026 · 5.50% · Board snapshotElevenLabs Scribe v2 · 5.93% · Board snapshotMusci-ASR · 5.70% · Board snapshotMusci-ASRSOV-Transcribe-1 · 5.96% · Board snapshotSOV-1Qwen3-ASR-1.7B · 6.15% · Board snapshotCohere Transcribe · 6.72% · Board snapshotCanary-Qwen 2.5B · 6.92% · Board snapshotParakeet TDT v2 · 6.93% · Board snapshotTDT v2Parakeet TDT v3 · 7.04% · Board snapshotVoxtral Small · 7.44% · Board snapshotVoxtral SmallWhisper Large v3 · 8.24% · Board snapshotVoxtral Mini · 8.32% · Board snapshotWhisper Large v3 Turbo · 8.75% · Board snapshotWhisper Large v3 TurboOruk v1 · 5.97% · Board snapshotOruk v1Oruk v1 · 5.97% · Board snapshotoruk · Open ASR v1.16.2 · 2026-11-20
Board result☆ Oruk board result

Full board · 7 public + 2 private · 1.7B nominal size. View evidence ↗

Board snapshot November 20, 2026. Source board ↗

Results by evaluation

Read the methodology →

The default English board averages seven public sets and two private tracks. Only published board scores appear in this table.

Open ASR · full results. Lower WER is better. Select a column to sort.
Zoom Scribe v2.1 ProClosed API4.87%Board snapshot ↗Full board · 7 public + 2 private
Otter Speech v3.1Closed API5.21%Board snapshot ↗Full board · 7 public + 2 private
Azure Speech 09-2026Closed API5.50%Board snapshot ↗Full board · 7 public + 2 private
Musci-ASR1.7B5.70%Board snapshot ↗Full board · 7 public + 2 private
ElevenLabs Scribe v2Closed API5.93%Board snapshot ↗Full board · 7 public + 2 private
SOV-Transcribe-126.6B5.96%Board snapshot ↗Full board · 7 public + 2 private
Oruk v11.7B5.97%Board snapshot ↗Full board · 7 public + 2 private
Qwen3-ASR-1.7B1.7B6.15%Board snapshot ↗Full board · 7 public + 2 private
Cohere Transcribe2B6.72%Board snapshot ↗Full board · 7 public + 2 private
Canary-Qwen 2.5B2.5B6.92%Board snapshot ↗Full board · 7 public + 2 private
Parakeet TDT v20.6B6.93%Board snapshot ↗Full board · 7 public + 2 private
Parakeet TDT v30.6B7.04%Board snapshot ↗Full board · 7 public + 2 private
Voxtral Small24B7.44%Board snapshot ↗Full board · 7 public + 2 private
Whisper Large v31.55B8.24%Board snapshot ↗Full board · 7 public + 2 private
Voxtral Mini3B8.32%Board snapshot ↗Full board · 7 public + 2 private
Whisper Large v3 Turbo0.81B8.75%Board snapshot ↗Full board · 7 public + 2 private

Roadmap targets

Campaign 4 defines separate goals for each leaderboard. Official scores require a published board result.

  • Oruk-1.7B · 4.95% WER target · Full board · 7 public + 2 private
  • Oruk-1.7B · 1.95% WER target · AA-WER v2.2 · non-streaming

Internal estimates

  • Oruk S1 · 5.52% WER · Internal projection · Full board · 7 public + 2 private
  • Qwen3-ASR-1.7B · 2.84% WER · Proxy estimate · Clean-9 calibration · internal
  • Oruk v2-rc3 · 2.40% WER · Proxy estimate · Clean-9 calibration · internal